This is the sixth blog post of Object Detection with YOLO blog series. This blog performs inference using the model in trained in Part 5 Object Detection with Yolo using VOC 2012 data - training. I will use PASCAL VOC2012 data. This blog assumes that the readers have read the previous blog posts - Part 1, Part 2, Part 3, Part 4, Part 5.
Andrew Ng's YOLO lecture¶
- Neural Networks - Bounding Box Predictions
- C4W3L06 Intersection Over Union
- C4W3L07 Nonmax Suppression
- C4W3L08 Anchor Boxes
- C4W3L09 YOLO Algorithm
Reference¶
Reference in my blog¶
- Part 1 Object Detection using YOLOv2 on Pascal VOC2012 - anchor box clustering
- Part 2 Object Detection using YOLOv2 on Pascal VOC2012 - input and output encoding
- Part 3 Object Detection using YOLOv2 on Pascal VOC2012 - model
- Part 4 Object Detection using YOLOv2 on Pascal VOC2012 - loss
- Part 5 Object Detection using YOLOv2 on Pascal VOC2012 - training
- Part 6 Object Detection using YOLOv2 on Pascal VOC 2012 data - inference on image
- Part 7 Object Detection using YOLOv2 on Pascal VOC 2012 data - inference on video
My GitHub repository¶
This repository contains all the ipython notebooks in this blog series and the funcitons (See backend.py).
In [1]:
import matplotlib.pyplot as plt
import numpy as np
import os, sys
print(sys.version)
%matplotlib inline
Read in the hyperparameters to define the YOLOv2 model used during training
In [2]:
train_image_folder = "../ObjectDetectionRCNN/VOCdevkit/VOC2012/JPEGImages/"
train_annot_folder = "../ObjectDetectionRCNN/VOCdevkit/VOC2012/Annotations/"
LABELS = ['aeroplane', 'bicycle', 'bird', 'boat', 'bottle',
'bus', 'car', 'cat', 'chair', 'cow',
'diningtable','dog', 'horse', 'motorbike', 'person',
'pottedplant','sheep', 'sofa', 'train', 'tvmonitor']
ANCHORS = np.array([1.07709888, 1.78171903, # anchor box 1, width , height
2.71054693, 5.12469308, # anchor box 2, width, height
10.47181473, 10.09646365, # anchor box 3, width, height
5.48531347, 8.11011331]) # anchor box 4, width, height
BOX = int(len(ANCHORS)/2)
TRUE_BOX_BUFFER = 50
IMAGE_H, IMAGE_W = 416, 416
GRID_H, GRID_W = 13 , 13
Load the weights trained in Part 5
In [3]:
from backend import define_YOLOv2
CLASS = len(LABELS)
model, _ = define_YOLOv2(IMAGE_H,IMAGE_W,GRID_H,GRID_W,TRUE_BOX_BUFFER,BOX,CLASS,
trainable=False)
model.load_weights("weights_yumi.h5")
In [4]:
## input encoding
from backend import ImageReader
imageReader = ImageReader(IMAGE_H,IMAGE_W=IMAGE_W, norm=lambda image : image / 255.)
out = imageReader.fit(train_image_folder + "/2007_005430.jpg")
Predict the bounding box.¶
In [5]:
print(out.shape)
X_test = np.expand_dims(out,0)
print(X_test.shape)
# handle the hack input
dummy_array = np.zeros((1,1,1,1,TRUE_BOX_BUFFER,4))
y_pred = model.predict([X_test,dummy_array])
print(y_pred.shape)
Rescale the network output¶
Remind you that y_pred takes any real values.
Therefore
In [6]:
class OutputRescaler(object):
def __init__(self,ANCHORS):
self.ANCHORS = ANCHORS
def _sigmoid(self, x):
return 1. / (1. + np.exp(-x))
def _softmax(self, x, axis=-1, t=-100.):
x = x - np.max(x)
if np.min(x) < t:
x = x/np.min(x)*t
e_x = np.exp(x)
return e_x / e_x.sum(axis, keepdims=True)
def get_shifting_matrix(self,netout):
GRID_H, GRID_W, BOX = netout.shape[:3]
no = netout[...,0]
ANCHORSw = self.ANCHORS[::2]
ANCHORSh = self.ANCHORS[1::2]
mat_GRID_W = np.zeros_like(no)
for igrid_w in range(GRID_W):
mat_GRID_W[:,igrid_w,:] = igrid_w
mat_GRID_H = np.zeros_like(no)
for igrid_h in range(GRID_H):
mat_GRID_H[igrid_h,:,:] = igrid_h
mat_ANCHOR_W = np.zeros_like(no)
for ianchor in range(BOX):
mat_ANCHOR_W[:,:,ianchor] = ANCHORSw[ianchor]
mat_ANCHOR_H = np.zeros_like(no)
for ianchor in range(BOX):
mat_ANCHOR_H[:,:,ianchor] = ANCHORSh[ianchor]
return(mat_GRID_W,mat_GRID_H,mat_ANCHOR_W,mat_ANCHOR_H)
def fit(self, netout):
'''
netout : np.array of shape (N grid h, N grid w, N anchor, 4 + 1 + N class)
a single image output of model.predict()
'''
GRID_H, GRID_W, BOX = netout.shape[:3]
(mat_GRID_W,
mat_GRID_H,
mat_ANCHOR_W,
mat_ANCHOR_H) = self.get_shifting_matrix(netout)
# bounding box parameters
netout[..., 0] = (self._sigmoid(netout[..., 0]) + mat_GRID_W)/GRID_W # x unit: range between 0 and 1
netout[..., 1] = (self._sigmoid(netout[..., 1]) + mat_GRID_H)/GRID_H # y unit: range between 0 and 1
netout[..., 2] = (np.exp(netout[..., 2]) + mat_ANCHOR_W)/GRID_W # width unit: range between 0 and 1
netout[..., 3] = (np.exp(netout[..., 3]) + mat_ANCHOR_H)/GRID_H # height unit: range between 0 and 1
# rescale the confidence to range 0 and 1
netout[..., 4] = self._sigmoid(netout[..., 4])
expand_conf = np.expand_dims(netout[...,4],-1) # (N grid h , N grid w, N anchor , 1)
# rescale the class probability to range between 0 and 1
# Pr(object class = k) = Pr(object exists) * Pr(object class = k |object exists)
# = Conf * P^c
netout[..., 5:] = expand_conf * self._softmax(netout[..., 5:])
# ignore the class probability if it is less than obj_threshold
return(netout)
Experiment OutputRescaler¶
In [7]:
netout = y_pred[0]
outputRescaler = OutputRescaler(ANCHORS=ANCHORS)
netout_scale = outputRescaler.fit(netout)
Post processing the YOLOv2 object¶
YOLOv2 can potentially preoduce GRID_H x GRID_W x BOX many bounding box. However, only few of them actually contain actual objects. Some bounding box may contain the same objects. I will postprocess the predicted bounding box.
In [8]:
from backend import BoundBox
def find_high_class_probability_bbox(netout_scale, obj_threshold):
'''
== Input ==
netout : y_pred[i] np.array of shape (GRID_H, GRID_W, BOX, 4 + 1 + N class)
x, w must be a unit of image width
y, h must be a unit of image height
c must be in between 0 and 1
p^c must be in between 0 and 1
== Output ==
boxes : list containing bounding box with Pr(object is in class C) > 0 for at least in one class C
'''
GRID_H, GRID_W, BOX = netout_scale.shape[:3]
boxes = []
for row in range(GRID_H):
for col in range(GRID_W):
for b in range(BOX):
# from 4th element onwards are confidence and class classes
classes = netout_scale[row,col,b,5:]
if np.sum(classes) > 0:
# first 4 elements are x, y, w, and h
x, y, w, h = netout_scale[row,col,b,:4]
confidence = netout_scale[row,col,b,4]
box = BoundBox(x-w/2, y-h/2, x+w/2, y+h/2, confidence, classes)
if box.get_score() > obj_threshold:
boxes.append(box)
return(boxes)
Experiment find_high_class_probability_bbox¶
In [9]:
obj_threshold = 0.015
boxes_tiny_threshold = find_high_class_probability_bbox(netout_scale,obj_threshold)
print("obj_threshold={}".format(obj_threshold))
print("In total, YOLO can produce GRID_H * GRID_W * BOX = {} bounding boxes ".format( GRID_H * GRID_W * BOX))
print("I found {} bounding boxes with top class probability > {}".format(len(boxes_tiny_threshold),obj_threshold))
obj_threshold = 0.03
boxes = find_high_class_probability_bbox(netout_scale,obj_threshold)
print("\nobj_threshold={}".format(obj_threshold))
print("In total, YOLO can produce GRID_H * GRID_W * BOX = {} bounding boxes ".format( GRID_H * GRID_W * BOX))
print("I found {} bounding boxes with top class probability > {}".format(len(boxes),obj_threshold))
Visualize many bounding box by having small obj_threshold value¶
Most of the bounding boxes do not contain objects. This shows that we really need to reduce the number of bounding box.
In [10]:
import cv2, copy
def draw_boxes(image, boxes, labels, obj_baseline=0.05,verbose=False):
'''
image : np.array of shape (N height, N width, 3)
'''
def adjust_minmax(c,_max):
if c < 0:
c = 0
if c > _max:
c = _max
return c
image = copy.deepcopy(image)
image_h, image_w, _ = image.shape
score_rescaled = np.array([box.get_score() for box in boxes])
score_rescaled /= obj_baseline
for sr, box in zip(score_rescaled,boxes):
xmin = adjust_minmax(int(box.xmin*image_w),image_w)
ymin = adjust_minmax(int(box.ymin*image_h),image_h)
xmax = adjust_minmax(int(box.xmax*image_w),image_w)
ymax = adjust_minmax(int(box.ymax*image_h),image_h)
text = "{:10} {:4.3f}".format(labels[box.label], box.get_score())
if verbose:
print("{} xmin={:4.0f},ymin={:4.0f},xmax={:4.0f},ymax={:4.0f}".format(text,xmin,ymin,xmax,ymax,text))
cv2.rectangle(image,
pt1=(xmin,ymin),
pt2=(xmax,ymax),
color=(0,1,0),
thickness=sr)
cv2.putText(img = image,
text = text,
org = (xmin+ 13, ymin + 13),
fontFace = cv2.FONT_HERSHEY_SIMPLEX,
fontScale = 1e-3 * image_h,
color = (1, 0, 1),
thickness = 1)
return image
print("Plot with low object threshold")
ima = draw_boxes(X_test[0],boxes_tiny_threshold,LABELS,verbose=True)
figsize = (15,15)
plt.figure(figsize=figsize)
plt.imshow(ima);
plt.title("Plot with low object threshold")
plt.show()
print("Plot with high object threshold")
ima = draw_boxes(X_test[0],boxes,LABELS,verbose=True)
figsize = (15,15)
plt.figure(figsize=figsize)
plt.imshow(ima);
plt.title("Plot with high object threshold")
plt.show()
Nonmax surpression¶
Nonmax surpression is a way to detect a single object only once. Andrew Ng has presented the idea of nonmax supression in his lecture very well: C4W3L07 Nonmax Suppression.
The following code implement the nonmax surpression algorithm. For each object class, the algorithm picks the most promissing bounding box, and then remove (or suppress) the remaining bounding box that contain high overwrap with the most promissing bounding box. The most promissing or not is determined by the predicted class probaiblity.
In [11]:
from backend import BestAnchorBoxFinder
def nonmax_suppression(boxes,iou_threshold,obj_threshold):
'''
boxes : list containing "good" BoundBox of a frame
[BoundBox(),BoundBox(),...]
'''
bestAnchorBoxFinder = BestAnchorBoxFinder([])
CLASS = len(boxes[0].classes)
index_boxes = []
# suppress non-maximal boxes
for c in range(CLASS):
# extract class probabilities of the c^th class from multiple bbox
class_probability_from_bbxs = [box.classes[c] for box in boxes]
#sorted_indices[i] contains the i^th largest class probabilities
sorted_indices = list(reversed(np.argsort( class_probability_from_bbxs)))
for i in range(len(sorted_indices)):
index_i = sorted_indices[i]
# if class probability is zero then ignore
if boxes[index_i].classes[c] == 0:
continue
else:
index_boxes.append(index_i)
for j in range(i+1, len(sorted_indices)):
index_j = sorted_indices[j]
# check if the selected i^th bounding box has high IOU with any of the remaining bbox
# if so, the remaining bbox' class probabilities are set to 0.
bbox_iou = bestAnchorBoxFinder.bbox_iou(boxes[index_i], boxes[index_j])
if bbox_iou >= iou_threshold:
classes = boxes[index_j].classes
classes[c] = 0
boxes[index_j].set_class(classes)
newboxes = [ boxes[i] for i in index_boxes if boxes[i].get_score() > obj_threshold ]
return newboxes
Experiment nonmax_suppression¶
In [12]:
iou_threshold = 0.01
final_boxes = nonmax_suppression(boxes,iou_threshold=iou_threshold,obj_threshold=obj_threshold)
print("{} final number of boxes".format(len(final_boxes)))
Finally draw the bounding box on an wapred image¶
In [13]:
ima = draw_boxes(X_test[0],final_boxes,LABELS,verbose=True)
figsize = (15,15)
plt.figure(figsize=figsize)
plt.imshow(ima);
plt.show()
More examples¶
In [14]:
np.random.seed(1)
Nsample = 20
image_nms = list(np.random.choice(os.listdir(train_image_folder),Nsample))
In [15]:
outputRescaler = OutputRescaler(ANCHORS=ANCHORS)
imageReader = ImageReader(IMAGE_H,IMAGE_W=IMAGE_W, norm=lambda image : image / 255.)
X_test = []
for img_nm in image_nms:
_path = os.path.join(train_image_folder,img_nm)
out = imageReader.fit(_path)
X_test.append(out)
X_test = np.array(X_test)
## model
dummy_array = np.zeros((len(X_test),1,1,1,TRUE_BOX_BUFFER,4))
y_pred = model.predict([X_test,dummy_array])
for iframe in range(len(y_pred)):
netout = y_pred[iframe]
netout_scale = outputRescaler.fit(netout)
boxes = find_high_class_probability_bbox(netout_scale,obj_threshold)
if len(boxes) > 0:
final_boxes = nonmax_suppression(boxes,
iou_threshold=iou_threshold,
obj_threshold=obj_threshold)
ima = draw_boxes(X_test[iframe],final_boxes,LABELS,verbose=True)
plt.figure(figsize=figsize)
plt.imshow(ima);
plt.show()
FairyOnIce/ObjectDetectionYolo contains this ipython notebook and all the functions that I defined in this notebook.